Scholarlink

<p><i>Objective:</i> Diabetic retinopathy (DR) is one of the leading causes of blindness. However, due to insufficient data and the computational complexity of existing deep learning models, capturing the characteristics of DR data remains challenging. Therefore, it is essential to develop a model that can effectively balance the detailed features and global information of images, while maintaining moderate computational complexity for DR detection. <i>Approach:</i> We propose a deep learning model based on optical coherence tomography (OCT) images for DR detection. Specifically, we propose a Conv-attention module to extract the subtle lesion features from DR data while retaining the global information. A novel fully quantized convolutional vision transformer (FQ-Conv-ViT) model is developed based on the Conv-attention module to enhance the performance of DR classification. By introducing quantization layers, the computational complexity of the proposed model is significantly reduced, and with the designed Conv-attention module, the overall feature presentation ability for DR detection is enhanced. <i>Main results:</i> Extensive experiments on two public datasets are conducted to validate the effectiveness of the proposed method. On the DRAC2022 dataset, we achieve a quadratic weighted kappa of 0.8074 and an accuracy of 85.94%. On the APTOS2019 dataset, we achieve a quadratic weighted kappa of 0.8933 and an accuracy of 87.97%. <i>Significance:</i> Our method outperforms existing state-of-the-art methods, demonstrates superior feature extraction capabilities and faster model inference speed, and can improve the overall performance of DR detection on OCT images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FQ-Conv-ViT: A quantized convolutional vision transformer model for diabetic retinopathy classification

  • 期刊论文

  • 发表日期：

    2025年5月31日

查看全文
  • Panpan Wu,
  • Yue Qu,
  • Ziping Zhao,
  • Zhangda Liu,
  • Hengyong Yu

摘要

Objective: Diabetic retinopathy (DR) is one of the leading causes of blindness. However, due to insufficient data and the computational complexity of existing deep learning models, capturing the characteristics of DR data remains challenging. Therefore, it is essential to develop a model that can effectively balance the detailed features and global information of images, while maintaining moderate computational complexity for DR detection. Approach: We propose a deep learning model based on optical coherence tomography (OCT) images for DR detection. Specifically, we propose a Conv-attention module to extract the subtle lesion features from DR data while retaining the global information. A novel fully quantized convolutional vision transformer (FQ-Conv-ViT) model is developed based on the Conv-attention module to enhance the performance of DR classification. By introducing quantization layers, the computational complexity of the proposed model is significantly reduced, and with the designed Conv-attention module, the overall feature presentation ability for DR detection is enhanced. Main results: Extensive experiments on two public datasets are conducted to validate the effectiveness of the proposed method. On the DRAC2022 dataset, we achieve a quadratic weighted kappa of 0.8074 and an accuracy of 85.94%. On the APTOS2019 dataset, we achieve a quadratic weighted kappa of 0.8933 and an accuracy of 87.97%. Significance: Our method outperforms existing state-of-the-art methods, demonstrates superior feature extraction capabilities and faster model inference speed, and can improve the overall performance of DR detection on OCT images.